Few-shot segmentation with duplex network and attention augmented module.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot segmentation faces challenges in diverse scenarios due to limited samples and insufficient exploration of support-query interactions.
• Overlooked interactions can cause model failures, especially with ambiguous boundaries in complex environments.

Purpose of the Study:

• To propose a novel network, Dynamic Prototype Mixture Convolutional Networks (DPMC), addressing the limitations of existing few-shot segmentation models.
• To enhance the focus on foreground objects by suppressing background noise and improving support-query interaction.

Main Methods:

• Developed a duplex network incorporating dynamic convolution for improved support-query interaction.
• Introduced a prototype match structure for comprehensive information extraction from support and query sets.
• Integrated a hybrid attentional module, double-layer attention augmented convolutional module (DAAConv), to minimize redundant information and enhance foreground focus.

Main Results:

• The proposed DPMC model, enhanced with DAAConv, demonstrated superior performance on PASCAL-5i and COCO-20i datasets.
• Achieved an average performance improvement of 5-8% over traditional prototype-based methods.
• Effectively suppresses background and focuses on foreground segmentation, even with ambiguous boundaries.

Conclusions:

• DPMC and DAAConv offer a significant advancement in few-shot segmentation by effectively handling limited data and complex scenarios.
• The proposed methods provide a robust solution for improving segmentation accuracy through enhanced feature interaction and attention mechanisms.
• Future research can explore further refinements of attention mechanisms and dynamic convolution for even greater segmentation performance.